13.2. THE CONDITION NUMBER 333

and so ap = bq at the point where this function has a minimum. Thus at this value of a,

f (a) =1

pap +

1

qap − aap−1 = ap − ap = 0

Hence f (a) ≥ 0 for all a ≥ 0 and this proves the inequality. Equality occurs when ap = bq.■

Now ||A||p may be considered as the operator norm of A taken with respect to ||·||p . Inthe case when p = 2, this is just the spectral norm. There is an easy estimate for ||A||p interms of the entries of A.

Theorem 13.1.5 The following holds.

||A||p ≤

∑k

∑j

|Ajk|pq/p

1/q

Proof: Let ||x||p ≤ 1 and let A = (a1, · · · ,an) where the ak are the columns of A. Then

Ax =

(∑k

xkak

)

and so by Holder’s inequality,

||Ax||p ≡

∣∣∣∣∣∣∣∣∣∣∑

k

xkak

∣∣∣∣∣∣∣∣∣∣p

≤∑k

|xk| ||ak||p ≤

(∑k

|xk|p)1/p(∑

k

||ak||qp

)1/q

∑k

∑j

|Ajk|pq/p

1/q

13.2 The Condition Number

Let A ∈ L (X,X) be a linear transformation where X is a finite dimensional vector spaceand consider the problem Ax = b where it is assumed there is a unique solution to thisproblem. How does the solution change if A is changed a little bit and if b is changed alittle bit? This is clearly an interesting question because you often do not know A and bexactly. If a small change in these quantities results in a large change in the solution, x,then it seems clear this would be undesirable. In what follows ||·|| when applied to a lineartransformation will always refer to the operator norm. Recall the following property of theoperator norm in Theorem 13.0.10.

Lemma 13.2.1 Let A,B ∈ L (X,X) where X is a normed vector space as above. Then for||·|| denoting the operator norm,

∥AB∥ ≤ ∥A∥ ∥B∥ .

Lemma 13.2.2 Let A,B ∈ L (X,X) , A−1 ∈ L (X,X) , and suppose ∥B∥ < 1/∥∥A−1

∥∥ .Then (A+B)

−1,(I +A−1B

)−1exists and∥∥∥(I +A−1B

)−1∥∥∥ ≤

(1−

∥∥A−1B∥∥)−1

(13.6)

13.2. THE CONDITION NUMBER 333and so a? = 6% at the point where this function has a minimum. Thus at this value of a,1, 1 4f (a) = —a? + —a? — aa? = aP —aP =0qdHence f (a) > 0 for all a > 0 and this proves the inequality. Equality occurs when a? = b!.aNow ||Al|,, may be considered as the operator norm of A taken with respect to ||-||,,. Inthe case when p = 2, this is just the spectral norm. There is an easy estimate for ||A|],, interms of the entries of A.Theorem 13.1.5 The following holds.a/p\ /4All, < | do | do Ase!k JProof: Let ||x||,, < 1 and let A = (a1,--- ,an) where the a; are the columns of A. Then(aand so by Holder’s inequality,|| Ax||,, |» ThAkk< S- |ezl llaall, <D ka/p\ 1/4<(Zhr) | (Sims) » Yl 7k13.2 The Condition NumberLet A € £(X,X) be a linear transformation where X is a finite dimensional vector spaceand consider the problem Ax = b where it is assumed there is a unique solution to thisproblem. How does the solution change if A is changed a little bit and if b is changed alittle bit? This is clearly an interesting question because you often do not know A and bexactly. If a small change in these quantities results in a large change in the solution, z,then it seems clear this would be undesirable. In what follows ||-|| when applied to a lineartransformation will always refer to the operator norm. Recall the following property of theoperator norm in Theorem 13.0.10.Lemma 13.2.1 Let A,B e £L(X,X) where X is a normed vector space as above. Then for||-|| denoting the operator norm,||AB|| < ||A}] BI.Lemma 13.2.2 Let A,B € £(X,X),A~! € L(X,X), and suppose ||B\| < 1/||A7*].Then (A+ B)', (I+ AB) exists and|(¢+4-tB) ||] < (1 atl) (13.6)